Mo-GAT: Multi-Objective Spatiotemporal Graph Attention Network for Dynamic Women Safety Risk Prediction and Proactive Prevention Routing

Abstract

Women\'s safety in Indian cities is a problem that has not improved as fast as anyone would like. In Andhra Pradesh, 16,863 cases of crimes against women were recorded in 2022 alone, and a large share of these incidents happened after dark in poorly lit, isolated, or overcrowded commercial areas. The mobile apps that exist today do almost nothing until after something goes wrong — they wait for the user to trigger an alert, and most of them stop working the moment network coverage drops, which is exactly when emergencies tend to happen. We built Mo-GAT to change that. It is an Android application that predicts danger before it materialises and still works when there is no signal at all. The underlying model is a two-layer Graph Attention Network trained on the MoGAT-Vijayawada-50K dataset — 50,000 geo-stamped records from 39 field-surveyed zones across Vijayawada, collected over twelve months. Each zone is described by six features: crime index, crowd density, street lighting quality, police station proximity, CCTV count, and a temporal risk factor. The network learns to assign a risk score to every zone and then uses those scores to recommend safer walking or travel routes, not just the shortest ones. On the test set, Mo-GAT reached 91.3% accuracy (MAE = 0.051, Pearson r = 0.937) and reduced average route risk by 34.7% against a standard shortest-path baseline. The SOS function fired within 1.8 seconds in full offline mode across 100 trials. A structured ablation study confirmed that both multi-head attention and the path-safety regulariser contribute independently to these gains.

Introduction

Urban mobility in Tier-2 Indian cities shows high safety risks, especially at night in poorly lit, low-police, and high-density pedestrian zones. Although crime statistics (e.g., NCRB data showing a 4% rise in crimes against women in 2022) highlight the issue, they fail to capture predictable spatial and temporal patterns of danger, such as high-risk zones varying by time, lighting, CCTV coverage, and police presence. Existing safety apps are largely reactive and depend heavily on internet connectivity, making them ineffective in high-risk areas where networks are often unreliable.

To address these limitations, the Mo-GAT system is proposed. It combines:

• An offline-first SOS system that can instantly trigger emergency calls, SMS alerts, and audio recording in under 1.8 seconds without internet.
• A spatiotemporal Graph Attention Network (GAT) trained on 50,000 geo-tagged records from 39 zones in Vijayawada to generate dynamic risk scores based on time and location.
• A proactive routing engine that guides users away from high-risk zones instead of only reacting after danger occurs.

The system is built on prior work in mobile safety apps, crime prediction models, and risk-aware routing, but improves upon them by integrating offline functionality, graph-based spatial learning, and joint optimization of prediction and routing.

A large dataset (MoGAT-Vijayawada-50K) was developed using real-world crime records, surveys, and infrastructure data, covering 39 zones and 21 features including crime index, crowd density, lighting, CCTV coverage, and police proximity. Expert validation ensured reliability of ground-truth risk scoring.

The system architecture includes a Flutter-based mobile app, Firebase backend, and a Python-based AI pipeline using PyTorch Geometric and TensorFlow Lite. It supports multiple SOS triggers (tap, volume buttons, shake) and ensures continuous emergency communication even without connectivity.

Conclusion

The design of Mo-GAT was motivated by two structural deficiencies identified across existing women\'s safety applications: reactive-only operation and dependence on continuous network connectivity. The proposed system addresses both deficiencies simultaneously. The risk prediction model — trained on 50,000 geo-stamped records from 39 Vijayawada zones — warns users away from danger before they enter it. The offline-first SOS engine responds in under two seconds with no network connection required. On the test set, the model hit 91.3% accuracy, produced zero critical misclassifications, and reduced average route risk by 34.7% against shortest-path navigation. Ablation results confirm that both the graph topology and the dual-objective loss contribute independently to those numbers; neither is a free ride from the other. Future work will pursue multi-city validation, real-time dynamic risk score updates from live sensor feeds, and federated learning to enable privacy-preserving model improvement from anonymised incident reports. The results presented here confirm that the individual components — offline SOS, graph-based risk prediction, and proactive routing — each function at the performance levels required for real-world deployment.

References

[1] National Crime Records Bureau (NCRB), \"Crime in India 2022,\" Ministry of Home Affairs, Government of India, New Delhi, 2023. [2] Andhra Pradesh State Crime Records Bureau, \"Annual Report on Crimes Against Women, Andhra Pradesh 2022,\" Amaravati: APSCRB, 2023. [3] N. Sharma, P. Malik and R. Agarwal, \"ABHAYA: An Android App for the Safety of Women,\" in Proc. IEEE ICCCA, 2015, pp. 1073–1077. [4] S. Mukherjee, R. Ghosh and A. Dey, \"SmartSafe: An Accelerometer-Based Women Safety System,\" in Proc. IEEE INDICON, 2018, pp. 1–6. [5] P. Srivastava, A. Misra and S. Kumar, \"Voice-Triggered Emergency SOS for Android Smartphones,\" Int. J. Embedded Systems, vol. 11, no. 4, pp. 422–431, 2019. [6] R. Patel and V. Sharma, \"A Systematic Review of Mobile Safety Applications for Women,\" IEEE Access, vol. 9, pp. 82140–82155, 2021. [7] L. Anselin, J. Cohen, D. Cook, W. Gorr and G. Tita, \"Spatial Analyses of Crime,\" Criminal Justice, vol. 4, pp. 213–262, 2000. [8] P. Wang, Y. Lin and X. Liu, \"Deep Spatio-Temporal Residual Networks for Citywide Crowd Flows Prediction,\" in Proc. AAAI, 2017, pp. 1655–1661. [9] B. Yu, H. Yin and Z. Zhu, \"Spatio-Temporal Graph Convolutional Networks,\" in Proc. IJCAI, 2018, pp. 3634–3640. [10] P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio and Y. Bengio, \"Graph Attention Networks,\" in Proc. ICLR, 2018. [11] J. Zhao, F. Chen and W. Dai, \"T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction,\" IEEE Trans. Intell. Transport. Syst., vol. 21, no. 9, pp. 3848–3858, 2020. [12] D. Quercia, R. Schifanella and L. M. Aiello, \"The Shortest Path to Happiness,\" in Proc. ACM HyperText, 2014, pp. 116–125. [13] P. Corcoran, C. Brunsdon and A. Ware, \"Modelling the Spatial Distribution of Urban Crime,\" Environment and Planning B, vol. 30, pp. 843–862, 2003. [14] TRAI, \"Telecom Subscription Data — December 2022,\" Telecom Regulatory Authority of India, New Delhi, Technical Report, 2023. [Online]. Available: https://www.trai.gov.in [15] S. Gayathri, K. U. Kumar, S. Bhuvaneshwari, V. Sathyapriya and C. Bhuvaneshwaran, \"Women Safety Device and Application — FEMME,\" in Proc. IEEE ICCSP, 2019, pp. 0937–0940. [16] K. Chaudhary, M. Yadav and B. Pradhan, \"A Review of Women Safety Services Using IoT and Machine Learning,\" in Proc. IEEE ICRITO, 2020, pp. 1–6. [17] H. Chen, Y. Li and F. Sun, \"Multi-Relational Graph Attention Networks for Urban Crime Prediction,\" in Proc. IEEE ICDM, 2021, pp. 1021–1026. [18] Y. Zhang, X. Wang and B. Liu, \"Incorporating Environmental Features into Spatiotemporal Crime Prediction,\" Expert Syst. Appl., vol. 195, art. 116523, 2022. [19] S. Brody, U. Alon and E. Yahav, \"How Attentive are Graph Attention Networks?,\" in Proc. ICLR, 2022. [20] K. Sila-Nowicka, J. Vandrol, T. Oshan, J. A. Long, U. Demšar and A. S. Fotheringham, \"Analysis of Human Mobility Patterns from GPS Trajectories and Contextual Information,\" Int. J. Geographical Inf. Sci., vol. 30, no. 5, pp. 881–906, 2016. [21] A. Hochmair, D. Zielstra and P. Neis, \"Assessing the Completeness of Bicycle Trail and Pedestrian Path Networks in OpenStreetMap for the Florida State,\" Trans. GIS, vol. 19, no. 1, pp. 82–103, 2015. [22] S. Ruan, C. Long and J. Zhang, \"Dynamic Safety-Aware Routing via Mobile Crowd Sensing,\" in Proc. ACM MobiCom, 2020, pp. 1–14. [23] T. N. Kipf and M. Welling, \"Semi-Supervised Classification with Graph Convolutional Networks,\" in Proc. ICLR, 2017. [24] J. R. Landis and G. G. Koch, \"The Measurement of Observer Agreement for Categorical Data,\" Biometrics, vol. 33, no. 1, pp. 159–174, Mar. 1977. [25] J. C. Nunnally, Psychometric Theory, 2nd ed. New York, NY, USA: McGraw-Hill, 1978. [26] A. Vaswani, N. Shazeer, N. Parmar et al., \"Attention Is All You Need,\" in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 30, 2017. [27] K. Starcke and M. Brand, \"Decision Making Under Stress: A Selective Review,\" Neuroscience & Biobehavioral Reviews, vol. 36, no. 4, pp. 1228–1248, 2012.

Copyright

Copyright © 2026 Damodara Harideviprathap, Borugadda Ravi Kiran, Sk. Badar , Sevugarajan , Dondapati Veera Swamy. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.